Manufacturing SystemsERP Platform

Databricks + ERP + AI

AI for Databricks in Manufacturing: Lakehouse Data From ERP, MES and the Line

Short answer

Databricks AI for manufacturing uses a lakehouse that holds ERP, MES and sensor data under Unity Catalog, then answers questions in plain English with the generated SQL shown. Genie spaces cover this inside Databricks; a private layer applies where models must run outside a public cloud or data cannot land there.

ERP: Databricks
Industries: Manufacturing, Aerospace, Electronics
Written for: Data Platform Lead

Discuss your Databricks environment See use cases

Databricks tends to enter a manufacturer through the data science door. A team wants to predict machine failures from vibration data, forecast demand beyond what the ERP's planning module does, or combine process data with quality results. Sensor streams, historian exports and MES events land as Delta tables, and ERP extracts join them so that a lot number in the quality system matches a purchase receipt in the ERP.

That combination is exactly what manufacturers cannot easily get elsewhere, and it is valuable. The difficulty is access. Notebooks and SQL warehouses serve data engineers, not planners or quality managers. The tables are big, the schemas reflect source systems rather than business concepts, and ERP columns need translating before anyone can ask a sensible question.

Databricks offers Genie, which answers questions over curated tables with instructions and sample queries, and Mosaic AI Model Serving for hosting models. For teams already on the platform and allowed to use it, these are the natural first step. Databricks is a cloud platform on AWS, Azure or GCP, with compute in the customer's cloud account, which is not enough for every air-gapped or export-controlled plant.

This page covers where AI on Databricks is useful for manufacturers, what governance and data modelling it needs, how it differs from classic BI, and when an on-prem alternative or simply a better-modelled warehouse is the honest recommendation.

What usually gets in the way

The problems we hear most from data platform lead teams running Databricks.

Schemas follow the source system, not the business

Bronze tables mirror ERP and MES structures with cryptic columns. Without curated silver and gold layers, questions return confident but wrong answers.

Time-series and transactional data do not align by default

Joining machine events with ERP work orders needs clock alignment, work center mapping and lot genealogy, which is real engineering.

Governance gaps between teams

Data science notebooks, BI users and AI assistants each reach data differently. Unity Catalog helps only when grants, masks and lineage are actually maintained.

Compute cost for exploratory use

Open-ended natural-language access on large Delta tables can run up serverless or cluster costs without limits and sensible table design.

Cloud platform, restricted data

Controlled technical data and some defense records may not be permitted on a public-cloud lakehouse regardless of the customer's own account ownership.

Where AI earns its place in Databricks

Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.

Natural-language questions over curated gold tables

Users ask about scrap by routing, supplier delivery performance or backlog. The layer maps to curated tables, generates SQL, and shows the query with the result.

Touches: Gold tables from ERP work orders, receipts, sales orders and BOMs

Outcome: Reduces ad hoc data-team requests for questions the gold layer already supports.

Machine event to ERP order correlation

A quality engineer asks which orders ran on a machine during a drift window. The layer joins MES or historian events with ERP work orders and lots.

Touches: Historian or MES event tables, ERP work orders, lot genealogy

Outcome: Faster scoping of containment, with the join logic visible for verification.

Predictive maintenance explanation

When a model flags an asset, the assistant summarizes recent sensor trends, work orders and parts history so the maintenance planner sees the context.

Touches: Sensor tables, CMMS or ERP maintenance work orders, spare parts consumption

Outcome: Predictions reach planners with context instead of a bare score.

Table and column documentation

The assistant profiles bronze and silver tables, proposes comments and descriptions in Unity Catalog, and flags undocumented columns for owners to confirm.

Touches: Unity Catalog metadata, column samples, lineage

Outcome: Better metadata for both humans and any model reading the catalog.

Pipeline failure triage

The assistant reads Lakeflow or Delta Live Tables event logs and job runs to explain failures and data freshness gaps.

Touches: Pipeline event logs, job run history, Auto Loader state

Outcome: Shorter path from stale data to a known cause.

Demand and inventory question support

Planners ask why a forecast moved, with the layer pulling the model's inputs, recent orders and inventory position into one explanation.

Touches: Forecast tables, sales history, inventory snapshots

Outcome: Forecasts become explainable, which helps planners trust or override them.

Document and data answers

Quality procedures, supplier contracts and lake tables are combined so answers cite both the numbers and the governing document.

Touches: Supplier and quality tables, contract and procedure documents

Outcome: Answers carry both the data and the rule it should be judged against.

Reference architecture

Inside Databricks, Genie works on curated tables under Unity Catalog. A private layer connects through a read-only SQL warehouse or the SQL connector, reuses the same curated definitions, and serves its model on infrastructure the customer controls.

1
Lakehouse connector
Connects through the Databricks SQL connector or SQL Statement API with a service principal mapped to the asking user's grants, so Unity Catalog permissions, row filters and column masks apply.
2
Curated layer
Silver and gold Delta tables with business names, comments, constraints and agreed joins, built from ERP, MES and sensor sources.
3
Model serving
Either Mosaic AI Model Serving inside Databricks, or an open-weight model such as Llama, Qwen or gpt-oss served with vLLM on customer GPUs, with only query results crossing the boundary.
4
Retrieval
Vector search over documents inside Databricks or a private index, combined with SQL answers when a question needs both.
5
Governance and audit
Query history, lineage and per-user logs show what was asked and what was read, with approval for any non-read action.

Integration notes for your ERP team

ERP data typically lands through Auto Loader, Lakeflow Connect, partner tools such as Fivetran, or SAP Datasphere exports into Delta tables.
Sensor and historian data arrives via streaming ingestion from OPC UA gateways, MQTT brokers or historian exports into bronze tables.
Build silver and gold layers with business names and column comments before pointing any assistant at the data.
Genie quality depends on curated tables, instructions, sample SQL and certified queries; budget time for this.
Use a small dedicated SQL warehouse with auto-stop and query timeouts for AI access to control spend.
Lot genealogy and work center mapping are required to join machine events to ERP orders correctly.
Record freshness per table and display it with each answer.

Deployment options

Inside Databricks with Genie

Teams already on Databricks whose policy allows it and whose gold tables are in good shape.

Genie spaces with instructions and sample queries over curated tables. Often the right first step.

Private model, lakehouse as source

Companies that need the language model outside a hosted service.

A read-only connector pulls masked query results to a model on customer GPUs or a private cloud, with a documented record of what is sent.

On-prem data platform

Plants whose restricted data cannot reside in a public-cloud lakehouse.

The same curated-layer approach on an on-prem database or Spark environment, with Databricks used only for non-restricted workloads.

Compliance and data control

How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.

Unity Catalog grants, row filters and column masks

AI queries run under mapped identities so the same access controls apply as for any SQL user.

ITAR and EAR technical data

If controlled data may not reside on a public-cloud lakehouse, it stays in an environment the customer controls; classification and placement remain the customer's decisions.

Data residency

Workspace region selection plus a private model deployment help meet residency requirements the customer has set.

Audit trails

Query history and lineage let quality and finance trace any cited figure back to its source tables.

Where Netray fits

DataRay

Chat with lakehouse tables, documents and file shares in one place, on-prem or in a private cloud, with the SQL shown.

Custom build

Predictive maintenance explanation, ERP and MES correlation and agents across systems are typically bespoke work.

How an engagement runs

Phase 1 . 2-4 weeks

Discovery

-Review of ingestion, layers and Unity Catalog setup
-Data classification and placement check
-Ranked question list
-Genie versus private layer recommendation

Phase 2 . 6-8 weeks

Pilot

-Curated tables for one domain
-Read-only access and masking review
-3-5 question patterns live
-Accuracy and cost review

Phase 3 . 4-6 weeks

Production

-Identity mapping and audit logging
-Cost guardrails
-Runbook for extending curated tables
-Handoff to the data team

Phase 4 . Ongoing

Scale

-More domains and plants
-Document retrieval
-Quarterly accuracy review
-Model refresh

Questions to ask any vendor, including us

A short list that separates real Databricks AI work from a chatbot demo.

Do Unity Catalog row filters and column masks apply to every AI query?
Are our gold tables ready, or is data modelling the real first project?
When is Genie enough and when would we need more?
How do you join machine events to ERP orders reliably?
How are compute costs capped for open-ended questions?
Is restricted data allowed on this platform at all?
Where does the model run and what leaves Databricks?
What do we keep if we stop?

Frequently asked questions

Is Genie enough for us?

If your curated tables are solid, your policy allows Databricks-hosted AI, and questions stay within the lakehouse, Genie may be enough. A private layer adds models outside a hosted service, restricted-data handling and cross-source or document questions.

What should be done before adding AI?

Build curated silver and gold layers with business names, comments and agreed joins. Models answer well over clean business tables and poorly over raw ERP mirrors. Expect the data modelling to take more effort than the AI itself.

Can it link sensor data with ERP orders?

Yes, if the mapping exists: machine to work center, timestamps aligned, and lot genealogy captured. We treat that mapping as a deliverable, and the join logic is shown with each answer so engineers can verify it.

Is this a replacement for our data science work?

No. Forecasting and predictive models stay with your data science team. The layer helps people ask questions of the results and the data behind them, and explains model outputs with context.

How is access controlled?

Through Unity Catalog. Queries run under mapped identities so grants, row filters and masks apply. We test with users in different roles and compare results to what each can see in SQL directly.

Can controlled data live on Databricks?

That is for your compliance team to decide. Databricks runs on public cloud infrastructure, even in your own account. If controlled data cannot be there, keep it in an environment you control and use the lakehouse for the rest.

Can the AI change ERP data?

No direct writes. It can draft recommendations or messages. Anything that changes ERP records goes through the ERP's workflow with a person approving it.

Related guides

Ask your ERP anything

Natural Language Query for ERP Data: Ask SAP, Infor, or Oracle a Question in Plain English

See how natural language query over SAP, Infor, Oracle, and NetSuite data works: grounded text-to-SQL, role-based permissions, and a visible audit trail.

RAG + SQL + permissions

A Private LLM Grounded on Your ERP Data

How a private LLM answers questions on your ERP data: RAG plus text-to-SQL, role-based permissions inherited from the ERP, and where each fits.

Snowflake + ERP + AI

AI for Snowflake in Manufacturing: Questions Across ERP, MES and Quality Data

Put grounded natural-language AI over ERP, MES and quality data in Snowflake, with role-based access honored, and see where Cortex fits versus a private model.

Agents + approval gates

AI Agents for ERP, Running On-Prem

A practical guide to on-prem AI agents for ERP: what they can safely automate, where human approval belongs, and how to design the guardrails.

ERP migration + on-prem AI

AI-assisted data migration for ERP implementations

AI speeds ERP data migration by profiling legacy data, proposing field mappings, and flagging cleansing issues before cutover, with a human validating every rule.

ERP AI readiness

Is Your ERP Ready for AI? A Readiness Assessment Checklist

Is your ERP actually ready for AI? A practical checklist covering master data quality, access and permissions, GPU sizing, and governance before you fund a pilot.

Plan it with numbers

Free Tool

Talk it through with an engineer who knows Databricks

Bring one real question your team cannot answer from the ERP today. We will map the data path, the model, and where it runs, and tell you honestly if AI is the wrong tool for it.

Start the conversation See ERPray

All AI for ERP guides